[epapop id=’3958′]
[responsivevoice_button voice=”US English Male” buttontext=”Listen the text”]
Specialists from the division DeepMind in London have created a deep training machine that is able to “learn” information in the shortest possible time. Compared with the machines of the previous generation, it works 10 times faster. Now, to master the classic video games, the mechanism requires not 200 hours of play, but 20, while a person needs only 2 hours.
The main idea of the DeepMind approach is to “teach” the machine to “think” quickly, like people and animals. As you know, a person can solve the situation in two different ways. First, if the situation is known, then our brain has already formed its model, which it uses to determine how to behave better. In this case, a part of the brain called the prefrontal cortex is used. If the situation is not familiar, then the brain relies on another strategy, which involves a simpler approach to verifying and remembering information. In this case, the hippocampus is used, which consolidates the memory.
For deep training machines use layers of neural networks to find patterns in the data. When one layer “sees” the template, it sends a signal as information to the next layer, which looks for patterns in this signal. Therefore, when recognizing faces, the first layer can search for the edges of the image, the second layer of circular images, the third – triangular patterns and so on. The final stage of the work of all layers is that the face was seen and recognized.
There are many feedback systems that allow the system to “learn” by regulating internal parameters, such as the strength of bonds between layers. The parameters should change slowly, since cardinal changes in one layer can negatively affect the overall process of “learning”. That’s why deep neural networks need such a long workout.
By author Anastasiya Stepanova